Guillaume Duvernay

Active

Lookio

RAG assistants you can embed on a site or call from a workflow.

You build an assistant over your own documents, then use it wherever you need it: a widget on your site, an API call from an automation, or a public assistant other people can use.

Role
Solo: product, engineering, marketing
Year
2025 to now
Stack
n8n, Qdrant, Cohere rerank, hybrid BM25 and semantic search

Three versions of the architecture

It started as an agent with a think tool. Then it became two modes: a fast one that answers in a single call with an array of questions, and an advanced one that runs the thinking as a sub-workflow. Then deep queries moved out of the agent entirely, into a workflow with an optional second retrieval pass. The change that helped most was using a small model for the agentic loop and a large one only to write the final answer.

Retrieval work

Cohere reranking, hybrid keyword and semantic search, and a summary of each document stored next to its chunks. The unglamorous fixes mattered most. One of the biggest quality jumps came from removing the words "in the documents" from a prompt, because they were degrading similarity scores on every query.

Where it's going

A library of public assistants you can call programmatically, with revenue share for the people who build them. Closer to context as a service than to another chat window.